Modern networks are increasingly exposed to cyber threats, making Network Intrusion Detection Systems (NIDS) an important part of network security. Machine learning-based NIDS can improve detection of attack patterns, but centralized training requires sensitive data to be transferred to a central server. Split learning addresses this issue by dividing a neural network between a client and a server, allowing raw data to remain on the client side. However, this architecture requires frequent communication between the client and server during training, making system performance sensitive to network conditions. The aim of this thesis is to provide practical insight into how network conditions affect split learning-based NIDS by evaluating the trade-offs between detection accuracy, communication cost, and training time. An experimental evaluation is conducted in a controlled cloud environment using two virtual machines. The implemented system is trained on the UNSW-NB15 dataset using ANN and CNN models, while different cut layer placements, packet loss levels, and transport protocols are evaluated. The findings show that detection accuracy and F1-score remain relatively stable in several conditions, while communication cost and training time are affected by model architecture, cut layer placement, and transport protocol. The results show that split learning-based NIDS should be evaluated not only by detection accuracy, but also by communication cost, training time, and network behaviour.